Abstract

The production of high yields of viable cells, especially Mesenchymal stem cells (MSCs), is a crucial yet challenging aspect in the field of cell therapy (CT). While progress has been made, there is still a need for quick, non-destructive ways to check the quality of the cells being produced to enhance cell manufacturing process. In light of this, our study aims to develop an accurate, interpretable machine learning technique that relies solely on bright-field (BF) images for enhanced differentiation of MSCs under different serum-containing conditions. Our investigation centers around the expansion of human MSCs derived from bone marrow cultivated in two specific media types: serum-containing (SC) and low-serum containing (LSC) media. The prevalent method of chemical staining for cell component identification is often time-intensive, costly, and potentially harmful to cells. To address these issues, we captured BF images at a 20X magnification with a Perkin Elmer Operatta screening system. Utilizing mean Shapley Adaptive exPlanations (SHAP) values obtained from the application of the 2-D discrete Fourier transform (DFT) module to BF images, we developed a supervised clustering approach within a tree-based machine learning model. The results of our experimental trials revealed the Random Forest model's efficacy in correctly classifying MSCs under varying conditions with a weighted accuracy of 80.15%. A further application of the DFT module to BF images significantly increased this accuracy to 93.26%. By transforming the original dataset into SHAP values using Random Forest classifiers, our supervised clustering approach effectively differentiates MSCs using label-free images. This innovative framework significantly contributes to the understanding of MSC health, enhances CT manufacturing processes, and holds potential to improve the efficacy of cell therapies.

© 2023 SPIE